Research On Manipulator Grasp Based On Deep Reinforcement Learning

The application of deep reinforcement learning methods to the control of robotic arms is currently a major research hotspot in the field of robotics.The traditional robotic arm control method has a strong dependence on the application scenarios of the robotic arm,and it is difficult to play a practical role in the unknown and changeable unstructured scene.Aiming at the many difficulties of robotic arm reinforcement learning grasping,this paper proposes a solution based on deep reinforcement learning.Based on deep learning,the target perception and pose estimation are combined with deep reinforcement learning to obtain the optimal control strategy.In addition,corresponding solutions are proposed for the simulationto-real problems in deep reinforcement learning.This article mainly conducts research from the following aspects:Firstly,for target detection and pose estimation tasks,a target detection network and a pose estimation network with a multi-layer backbone network are established respectively,and the network structure and parameters established are determined.The Label Me tool is used to establish a corresponding data set to train the target detection network,and the YCB data set is selected to train the pose estimation network.The target detection and pose estimation network established in this paper can obtain more accurate detection results in less training time.Secondly,the kinematics of Kinova's seven-degree-of-freedom manipulator is analyzed.Aiming at the problems of low exploration efficiency and long training time of deep reinforcement learning in the task training process of robotic arms,a deep reinforcement learning algorithm combining target detection and pose estimation is proposed.Perform target detection and pose estimation on the objects that need to be grasped,and then adjust the strategy in reinforcement learning according to the results obtained,and limit the range of motion and end posture of the robotic arm,thereby improving the exploration efficiency and finality of the training process.The success rate of crawling.This algorithm can effectively improve the efficiency of reinforcement learning training.Thirdly,the training process of deep reinforcement learning is generally completed in a simulation environment.Aiming at the problem of difficulty in migrating the strategies trained by deep reinforcement learning in the simulation environment to the real environment,a training based on domain randomization and fusion of real data and simulation data is proposed.method.The domain randomization method is used to establish a simulation environment with variable parameters,and the domain randomization strategy is updated with the data obtained from the application of the strategy in the real environment and simulation.Finally,the error of the model in the simulation environment is reduced,making it closer to the real environment.So that the strategies trained in the simulation environment can be directly applied to the real environment.In addition,a corresponding solution is proposed for the problem of sparse reward in deep reinforcement learning.Finally,a realistic experiment platform and a simulation training platform based on Coppelia Sim simulation software were built.In view of the above-mentioned research,relevant actual and simulation experiments are designed to verify the feasibility and superiority of the proposed scheme,and the experimental results are analyzed in detail.